rm

Performs pixel-level semantic segmentation to isolate foreground subjects from backgrounds using a transformer-based vision model trained on diverse image datasets. The model outputs binary or multi-class segmentation masks that can be directly applied to remove, replace, or isolate background regions. Works by processing images through a CNN-transformer hybrid architecture that captures both local spatial features and global context, enabling accurate boundary detection even with complex or blurred backgrounds.

Provides pre-exported model weights in multiple formats (PyTorch, ONNX, SafeTensors) enabling deployment across heterogeneous environments without retraining or conversion overhead. The model can be loaded directly via transformers library for Python, executed via ONNX Runtime for C++/C#/.NET/JavaScript environments, or imported into transformers.js for browser-based inference. This architecture decouples model definition from runtime, allowing the same trained weights to run on servers, edge devices, and client-side applications.

Supports processing multiple images sequentially or in batches through a standardized preprocessing pipeline that handles image resizing, normalization, and tensor conversion. The model accepts variable-resolution inputs and internally normalizes them to the training resolution using configurable interpolation methods (bilinear, nearest-neighbor). Preprocessing includes channel-wise normalization using ImageNet statistics, enabling consistent output quality across diverse image sources and lighting conditions.